Merge pull request #995 from FioraAeterna/fma

Add FMA support to emitter and use it in the JIT
This commit is contained in:
skidau 2014-10-12 13:56:18 +11:00
commit 9ddbdeb39f
6 changed files with 277 additions and 45 deletions

View File

@ -1355,9 +1355,9 @@ void XEmitter::WriteSSEOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extr
arg.WriteRest(this, extrabytes); arg.WriteRest(this, extrabytes);
} }
void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes) void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int W, int extrabytes)
{ {
WriteAVXOp(opPrefix, op, regOp, INVALID_REG, arg, extrabytes); WriteAVXOp(opPrefix, op, regOp, INVALID_REG, arg, W, extrabytes);
} }
static int GetVEXmmmmm(u16 op) static int GetVEXmmmmm(u16 op)
@ -1383,14 +1383,14 @@ static int GetVEXpp(u8 opPrefix)
return 0; return 0;
} }
void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes) void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int W, int extrabytes)
{ {
if (!cpu_info.bAVX) if (!cpu_info.bAVX)
PanicAlert("Trying to use AVX on a system that doesn't support it. Bad programmer."); PanicAlert("Trying to use AVX on a system that doesn't support it. Bad programmer.");
int mmmmm = GetVEXmmmmm(op); int mmmmm = GetVEXmmmmm(op);
int pp = GetVEXpp(opPrefix); int pp = GetVEXpp(opPrefix);
// FIXME: we currently don't support 256-bit instructions, and "size" is not the vector size here // FIXME: we currently don't support 256-bit instructions, and "size" is not the vector size here
arg.WriteVex(this, regOp1, regOp2, 0, pp, mmmmm); arg.WriteVex(this, regOp1, regOp2, 0, pp, mmmmm, W);
Write8(op & 0xFF); Write8(op & 0xFF);
arg.WriteRest(this, extrabytes, regOp1); arg.WriteRest(this, extrabytes, regOp1);
} }
@ -1799,10 +1799,71 @@ void XEmitter::VPAND(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x6
void XEmitter::VPANDN(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseANDN, regOp1, regOp2, arg);} void XEmitter::VPANDN(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseANDN, regOp1, regOp2, arg);}
void XEmitter::VPOR(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseOR, regOp1, regOp2, arg);} void XEmitter::VPOR(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseOR, regOp1, regOp2, arg);}
void XEmitter::VPXOR(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseXOR, regOp1, regOp2, arg);} void XEmitter::VPXOR(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, sseXOR, regOp1, regOp2, arg);}
void XEmitter::VSHUFPD(X64Reg regOp1, X64Reg regOp2, OpArg arg, u8 shuffle) {WriteAVXOp(0x66, sseSHUF, regOp1, regOp2, arg, 1); Write8(shuffle);} void XEmitter::VSHUFPD(X64Reg regOp1, X64Reg regOp2, OpArg arg, u8 shuffle) {WriteAVXOp(0x66, sseSHUF, regOp1, regOp2, arg, 0, 1); Write8(shuffle);}
void XEmitter::VUNPCKLPD(X64Reg regOp1, X64Reg regOp2, OpArg arg){WriteAVXOp(0x66, 0x14, regOp1, regOp2, arg);} void XEmitter::VUNPCKLPD(X64Reg regOp1, X64Reg regOp2, OpArg arg){WriteAVXOp(0x66, 0x14, regOp1, regOp2, arg);}
void XEmitter::VUNPCKHPD(X64Reg regOp1, X64Reg regOp2, OpArg arg){WriteAVXOp(0x66, 0x15, regOp1, regOp2, arg);} void XEmitter::VUNPCKHPD(X64Reg regOp1, X64Reg regOp2, OpArg arg){WriteAVXOp(0x66, 0x15, regOp1, regOp2, arg);}
void XEmitter::VFMADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x3898, regOp1, regOp2, arg);}
void XEmitter::VFMADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38A8, regOp1, regOp2, arg);}
void XEmitter::VFMADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38B8, regOp1, regOp2, arg);}
void XEmitter::VFMADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x3898, regOp1, regOp2, arg, 1);}
void XEmitter::VFMADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38A8, regOp1, regOp2, arg, 1);}
void XEmitter::VFMADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38B8, regOp1, regOp2, arg, 1);}
void XEmitter::VFMADD132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x3899, regOp1, regOp2, arg);}
void XEmitter::VFMADD213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38A9, regOp1, regOp2, arg);}
void XEmitter::VFMADD231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38B9, regOp1, regOp2, arg);}
void XEmitter::VFMADD132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x3899, regOp1, regOp2, arg, 1);}
void XEmitter::VFMADD213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38A9, regOp1, regOp2, arg, 1);}
void XEmitter::VFMADD231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38B9, regOp1, regOp2, arg, 1);}
void XEmitter::VFMSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389A, regOp1, regOp2, arg);}
void XEmitter::VFMSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AA, regOp1, regOp2, arg);}
void XEmitter::VFMSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BA, regOp1, regOp2, arg);}
void XEmitter::VFMSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389A, regOp1, regOp2, arg, 1);}
void XEmitter::VFMSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AA, regOp1, regOp2, arg, 1);}
void XEmitter::VFMSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BA, regOp1, regOp2, arg, 1);}
void XEmitter::VFMSUB132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389B, regOp1, regOp2, arg);}
void XEmitter::VFMSUB213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AB, regOp1, regOp2, arg);}
void XEmitter::VFMSUB231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BB, regOp1, regOp2, arg);}
void XEmitter::VFMSUB132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389B, regOp1, regOp2, arg, 1);}
void XEmitter::VFMSUB213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AB, regOp1, regOp2, arg, 1);}
void XEmitter::VFMSUB231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BB, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389C, regOp1, regOp2, arg);}
void XEmitter::VFNMADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AC, regOp1, regOp2, arg);}
void XEmitter::VFNMADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BC, regOp1, regOp2, arg);}
void XEmitter::VFNMADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389C, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AC, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BC, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMADD132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389D, regOp1, regOp2, arg);}
void XEmitter::VFNMADD213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AD, regOp1, regOp2, arg);}
void XEmitter::VFNMADD231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BD, regOp1, regOp2, arg);}
void XEmitter::VFNMADD132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389D, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMADD213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AD, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMADD231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BD, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389E, regOp1, regOp2, arg);}
void XEmitter::VFNMSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AE, regOp1, regOp2, arg);}
void XEmitter::VFNMSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BE, regOp1, regOp2, arg);}
void XEmitter::VFNMSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389E, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AE, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BE, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMSUB132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389F, regOp1, regOp2, arg);}
void XEmitter::VFNMSUB213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AF, regOp1, regOp2, arg);}
void XEmitter::VFNMSUB231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BF, regOp1, regOp2, arg);}
void XEmitter::VFNMSUB132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x389F, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMSUB213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38AF, regOp1, regOp2, arg, 1);}
void XEmitter::VFNMSUB231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38BF, regOp1, regOp2, arg, 1);}
void XEmitter::VFMADDSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x3896, regOp1, regOp2, arg);}
void XEmitter::VFMADDSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38A6, regOp1, regOp2, arg);}
void XEmitter::VFMADDSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38B6, regOp1, regOp2, arg);}
void XEmitter::VFMADDSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x3896, regOp1, regOp2, arg, 1);}
void XEmitter::VFMADDSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38A6, regOp1, regOp2, arg, 1);}
void XEmitter::VFMADDSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38B6, regOp1, regOp2, arg, 1);}
void XEmitter::VFMSUBADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x3897, regOp1, regOp2, arg);}
void XEmitter::VFMSUBADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38A7, regOp1, regOp2, arg);}
void XEmitter::VFMSUBADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38B7, regOp1, regOp2, arg);}
void XEmitter::VFMSUBADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x3897, regOp1, regOp2, arg, 1);}
void XEmitter::VFMSUBADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38A7, regOp1, regOp2, arg, 1);}
void XEmitter::VFMSUBADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg) {WriteAVXOp(0x66, 0x38B7, regOp1, regOp2, arg, 1);}
void XEmitter::SARX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0xF3, 0x38F7, regOp1, regOp2, arg);} void XEmitter::SARX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0xF3, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::SHLX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0x66, 0x38F7, regOp1, regOp2, arg);} void XEmitter::SHLX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0x66, 0x38F7, regOp1, regOp2, arg);}
void XEmitter::SHRX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0xF2, 0x38F7, regOp1, regOp2, arg);} void XEmitter::SHRX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2) {WriteBMI2Op(bits, 0xF2, 0x38F7, regOp1, regOp2, arg);}

View File

@ -294,8 +294,8 @@ private:
void WriteSSEOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0); void WriteSSEOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0);
void WriteSSSE3Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0); void WriteSSSE3Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0);
void WriteSSE41Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0); void WriteSSE41Op(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0);
void WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int extrabytes = 0); void WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp, OpArg arg, int W = 0, int extrabytes = 0);
void WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0); void WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int W = 0, int extrabytes = 0);
void WriteVEXOp(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0); void WriteVEXOp(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0);
void WriteBMI1Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0); void WriteBMI1Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0);
void WriteBMI2Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0); void WriteBMI2Op(int size, u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, OpArg arg, int extrabytes = 0);
@ -773,6 +773,68 @@ public:
void VUNPCKLPD(X64Reg regOp1, X64Reg regOp2, OpArg arg); void VUNPCKLPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VUNPCKHPD(X64Reg regOp1, X64Reg regOp2, OpArg arg); void VUNPCKHPD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
// FMA
void VFMADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADD231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUB231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMADD231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB132SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB213SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB231SS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB132SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB213SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFNMSUB231SD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMADDSUB231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD132PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD213PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD231PS(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD132PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD213PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
void VFMSUBADD231PD(X64Reg regOp1, X64Reg regOp2, OpArg arg);
// VEX GPR instructions // VEX GPR instructions
void SARX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2); void SARX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2);
void SHLX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2); void SHLX(int bits, X64Reg regOp1, OpArg arg, X64Reg regOp2);

View File

@ -49,6 +49,7 @@
#include "Core/HW/Wiimote.h" #include "Core/HW/Wiimote.h"
#include "Core/IPC_HLE/WII_IPC_HLE_Device_usb.h" #include "Core/IPC_HLE/WII_IPC_HLE_Device_usb.h"
#include "Core/IPC_HLE/WII_Socket.h" #include "Core/IPC_HLE/WII_Socket.h"
#include "Core/PowerPC/JitInterface.h"
#include "Core/PowerPC/PowerPC.h" #include "Core/PowerPC/PowerPC.h"
#ifdef USE_GDBSTUB #ifdef USE_GDBSTUB
@ -728,6 +729,8 @@ void UpdateWantDeterminism(bool initial)
g_want_determinism = new_want_determinism; g_want_determinism = new_want_determinism;
WiiSockMan::GetInstance().UpdateWantDeterminism(new_want_determinism); WiiSockMan::GetInstance().UpdateWantDeterminism(new_want_determinism);
g_video_backend->UpdateWantDeterminism(new_want_determinism); g_video_backend->UpdateWantDeterminism(new_want_determinism);
// We need to clear the cache because some parts of the JIT depend on want_determinism, e.g. use of FMA.
JitInterface::ClearCache();
Core::PauseAndLock(false, was_unpaused); Core::PauseAndLock(false, was_unpaused);
} }

View File

@ -90,9 +90,44 @@ void Jit64::fmaddXX(UGeckoInstruction inst)
fpr.Lock(a, b, c, d); fpr.Lock(a, b, c, d);
// nmsub is implemented a little differently ((b - a*c) instead of -(a*c - b)), so handle it separately // While we don't know if any games are actually affected (replays seem to work with all the usual
if (inst.SUBOP5 == 30) //nmsub // suspects for desyncing), netplay and other applications need absolute perfect determinism, so
// be extra careful and don't use FMA, even if in theory it might be okay.
// Note that FMA isn't necessarily less correct (it may actually be closer to correct) compared
// to what the Gekko does here; in deterministic mode, the important thing is multiple Dolphin
// instances on different computers giving identical results.
if (cpu_info.bFMA && !Core::g_want_determinism)
{ {
if (single_precision)
Force25BitPrecision(XMM0, fpr.R(c), XMM1);
else
MOVSD(XMM0, fpr.R(c));
// Statistics suggests b is a lot less likely to be unbound in practice, so
// if we have to pick one of a or b to bind, let's make it b.
fpr.BindToRegister(b, true, false);
switch (inst.SUBOP5)
{
case 28: //msub
VFMSUB132SD(XMM0, fpr.RX(b), fpr.R(a));
break;
case 29: //madd
VFMADD132SD(XMM0, fpr.RX(b), fpr.R(a));
break;
// PowerPC and x86 define NMADD/NMSUB differently
// x86: D = -A*C (+/-) B
// PPC: D = -(A*C (+/-) B)
// so we have to swap them; the ADD/SUB here isn't a typo.
case 30: //nmsub
VFNMADD132SD(XMM0, fpr.RX(b), fpr.R(a));
break;
case 31: //nmadd
VFNMSUB132SD(XMM0, fpr.RX(b), fpr.R(a));
break;
}
}
else if (inst.SUBOP5 == 30) //nmsub
{
// nmsub is implemented a little differently ((b - a*c) instead of -(a*c - b)), so handle it separately
if (single_precision) if (single_precision)
Force25BitPrecision(XMM1, fpr.R(c), XMM0); Force25BitPrecision(XMM1, fpr.R(c), XMM0);
else else
@ -115,6 +150,7 @@ void Jit64::fmaddXX(UGeckoInstruction inst)
if (inst.SUBOP5 == 31) //nmadd if (inst.SUBOP5 == 31) //nmadd
PXOR(XMM0, M((void*)&psSignBits)); PXOR(XMM0, M((void*)&psSignBits));
} }
fpr.BindToRegister(d, false); fpr.BindToRegister(d, false);
//YES it is necessary to dupe the result :( //YES it is necessary to dupe the result :(
//TODO : analysis - does the top reg get used? If so, dupe, if not, don't. //TODO : analysis - does the top reg get used? If so, dupe, if not, don't.

View File

@ -305,50 +305,77 @@ void Jit64::ps_maddXX(UGeckoInstruction inst)
int b = inst.FB; int b = inst.FB;
int c = inst.FC; int c = inst.FC;
int d = inst.FD; int d = inst.FD;
bool fma = cpu_info.bFMA && !Core::g_want_determinism;
fpr.Lock(a,b,c,d); fpr.Lock(a,b,c,d);
switch (inst.SUBOP5) if (fma)
fpr.BindToRegister(b, true, false);
if (inst.SUBOP5 == 14)
{ {
case 14: //madds0
MOVDDUP(XMM0, fpr.R(c)); MOVDDUP(XMM0, fpr.R(c));
Force25BitPrecision(XMM0, R(XMM0), XMM1); Force25BitPrecision(XMM0, R(XMM0), XMM1);
MULPD(XMM0, fpr.R(a)); }
ADDPD(XMM0, fpr.R(b)); else if (inst.SUBOP5 == 15)
break; {
case 15: //madds1
avx_op(&XEmitter::VSHUFPD, &XEmitter::SHUFPD, XMM0, fpr.R(c), fpr.R(c), 3); avx_op(&XEmitter::VSHUFPD, &XEmitter::SHUFPD, XMM0, fpr.R(c), fpr.R(c), 3);
Force25BitPrecision(XMM0, R(XMM0), XMM1); Force25BitPrecision(XMM0, R(XMM0), XMM1);
MULPD(XMM0, fpr.R(a));
ADDPD(XMM0, fpr.R(b));
break;
case 28: //msub
Force25BitPrecision(XMM0, fpr.R(c), XMM1);
MULPD(XMM0, fpr.R(a));
SUBPD(XMM0, fpr.R(b));
break;
case 29: //madd
Force25BitPrecision(XMM0, fpr.R(c), XMM1);
MULPD(XMM0, fpr.R(a));
ADDPD(XMM0, fpr.R(b));
break;
case 30: //nmsub
Force25BitPrecision(XMM0, fpr.R(c), XMM1);
MULPD(XMM0, fpr.R(a));
SUBPD(XMM0, fpr.R(b));
PXOR(XMM0, M((void*)&psSignBits));
break;
case 31: //nmadd
Force25BitPrecision(XMM0, fpr.R(c), XMM1);
MULPD(XMM0, fpr.R(a));
ADDPD(XMM0, fpr.R(b));
PXOR(XMM0, M((void*)&psSignBits));
break;
default:
_assert_msg_(DYNA_REC, 0, "ps_maddXX WTF!!!");
//FallBackToInterpreter(inst);
//fpr.UnlockAll();
return;
} }
else
{
Force25BitPrecision(XMM0, fpr.R(c), XMM1);
}
if (fma)
{
switch (inst.SUBOP5)
{
case 14: //madds0
case 15: //madds1
case 29: //madd
VFMADD132PD(XMM0, fpr.RX(b), fpr.R(a));
break;
case 28: //msub
VFMSUB132PD(XMM0, fpr.RX(b), fpr.R(a));
break;
case 30: //nmsub
VFNMADD132PD(XMM0, fpr.RX(b), fpr.R(a));
break;
case 31: //nmadd
VFNMSUB132PD(XMM0, fpr.RX(b), fpr.R(a));
break;
}
}
else
{
switch (inst.SUBOP5)
{
case 14: //madds0
case 15: //madds1
case 29: //madd
MULPD(XMM0, fpr.R(a));
ADDPD(XMM0, fpr.R(b));
break;
case 28: //msub
MULPD(XMM0, fpr.R(a));
SUBPD(XMM0, fpr.R(b));
break;
case 30: //nmsub
MULPD(XMM0, fpr.R(a));
SUBPD(XMM0, fpr.R(b));
PXOR(XMM0, M((void*)&psSignBits));
break;
case 31: //nmadd
MULPD(XMM0, fpr.R(a));
ADDPD(XMM0, fpr.R(b));
PXOR(XMM0, M((void*)&psSignBits));
break;
default:
_assert_msg_(DYNA_REC, 0, "ps_maddXX WTF!!!");
return;
}
}
fpr.BindToRegister(d, false); fpr.BindToRegister(d, false);
ForceSinglePrecisionP(fpr.RX(d), XMM0); ForceSinglePrecisionP(fpr.RX(d), XMM0);
SetFPRFIfNeeded(inst, fpr.RX(d)); SetFPRFIfNeeded(inst, fpr.RX(d));

View File

@ -948,4 +948,47 @@ VEX_RM_TEST(BLSI)
VEX_RMI_TEST(RORX) VEX_RMI_TEST(RORX)
// for AVX instructions that take the form op reg, reg, r/m
#define AVX_RRM_TEST(Name, sizename) \
TEST_F(x64EmitterTest, Name) \
{ \
struct { \
int bits; \
std::vector<NamedReg> regs; \
std::string out_name; \
std::string size; \
} regsets[] = { \
{ 64, xmmnames, "xmm0", sizename }, \
}; \
for (const auto& regset : regsets) \
for (const auto& r : regset.regs) \
{ \
emitter->Name(r.reg, RAX, R(RAX)); \
emitter->Name(RAX, RAX, R(r.reg)); \
emitter->Name(RAX, r.reg, MatR(R12)); \
ExpectDisassembly(#Name " " + r.name+ ", " + regset.out_name + ", " + regset.out_name + " " \
#Name " " + regset.out_name + ", " + regset.out_name + ", " + r.name + " " \
#Name " " + regset.out_name + ", " + r.name + ", " + regset.size + " ptr ds:[r12] "); \
} \
}
#define FMA_TEST(Name, P, packed) \
AVX_RRM_TEST(Name ## 132 ## P ## S, packed ? "dqword" : "dword") \
AVX_RRM_TEST(Name ## 213 ## P ## S, packed ? "dqword" : "dword") \
AVX_RRM_TEST(Name ## 231 ## P ## S, packed ? "dqword" : "dword") \
AVX_RRM_TEST(Name ## 132 ## P ## D, packed ? "dqword" : "qword") \
AVX_RRM_TEST(Name ## 213 ## P ## D, packed ? "dqword" : "qword") \
AVX_RRM_TEST(Name ## 231 ## P ## D, packed ? "dqword" : "qword")
FMA_TEST(VFMADD, P, true)
FMA_TEST(VFMADD, S, false)
FMA_TEST(VFMSUB, P, true)
FMA_TEST(VFMSUB, S, false)
FMA_TEST(VFNMADD, P, true)
FMA_TEST(VFNMADD, S, false)
FMA_TEST(VFNMSUB, P, true)
FMA_TEST(VFNMSUB, S, false)
FMA_TEST(VFMADDSUB, P, true)
FMA_TEST(VFMSUBADD, P, true)
} // namespace Gen } // namespace Gen